|
Author
|
Topic: Responses to Criticisms of Specified Complexity
|
Janitor@MIT
Member
Member # 125
|
posted 12. May 2003 15:26
I’ve had second thoughts and I apologize to Rex Kerr (and all the observers and participants) for being a shit. (Before Mr. Moderator requires me to do so.) If I am permitted to be “unscholarly,” and just “brainstorm,” I will (with Mr. Moderator’s tacit encouragement LOL) ramble on endlessly, irrelevantly, irreverently, and ultimately (possibly) pointlessly…
Early in the history of this board there were many interesting discussions about the application of information theory to biology. Many of the same questions that were asked and the same problems noticed then apply to the application of probability theory to biology. (Just translate P for H in Shannon’s theory and you can’t help but notice the near-identity. Shannon simply relates probability theory in the conventional language and to the basic problems of communications engineers.)
E.g., this argument (which I keep returning to) that selection somehow positively conditions variation is just a relaxation of the Markov condition. We’re not going to eliminate the Markov condition, because I doubt we could understand how probability w/o it is related to causality (Or how probability w/o it is probability!)—It’s just so central to our notions of causality and may be necessary for us to make any kind of sense of causation. In fact biologists already do this on an observational basis. They recognize that “events” are not narrowly proximally related.
Biologists also recognize that life processes are reflexive. (I will say that biologists don't really appreciate the problem here.)
Biologists also recognize that conditions of symmetry, transitivity, monotonicity, continuity, and simple additivity (or whatever you want to call it) do not strictly hold.
I have suggested (above) that if our notions of probability = strict causal dependency then we are missing something important about how genomes evolve by exploiting DOF (i.e., causal “independencies”).
Probably many such “exceptions” could be found. In other words, virtually every axiom of probability theory is questionable for biology.
Basically, what is required of us to erect a meaningful probabilistic model of evolution is adapting probability theoretic axioms to life, rather than other way round; life “adapting” to our axioms. We need to adapt our theories to the phenomenon. That is, after all, what science is about? Isn’t it?
Now I understand that Dr. Dembski is a probability theorist. (And I’m sure I just made a fool of myself in his eyes.) Regardless of how “unscholarly” I may appear, it is a frequent criticism of his thesis that it is not biologically “realistic.” But as I’ve attempted to say, it appears as if probability theory needs to be adapted to life. We need to discover how it is that life forms in their operations, their adaptation and evolution, do not abide by our notions of probability. How they have their own “notions” of probability. The IDers might consider it, as one of their, if not their central thesis, is that life is not “natural.” If the axioms of probability theory do in some sense define for us what is natural (and its helpful to consider the origins of probability theory and its close and continued association here with physics, and such confident assertions that the laws of physics are the laws of probability, and the laws of logic and science (i.e., inference) are probabilistic) then the “fact” that life forms “think otherwise” is interesting to consider.
My only familiarity with probability theory is through information and decision theory. One of the results of decision theory is also that a strong Markov condition cannot hold for decision (formally equivalent to evolution) processes. I.e., all that can be reasonably expected is that a decision-maker, a life form, achieves a partial and ad hoc ordering over preferences or utilities. Because for any interesting (LOL Mathematicians always define as "interesting" those problems that they cannot easily solve.)decision problem this strict ordering (which is just Markov) would take longer than the apparent age of the universe, i.e., exceed the computational capacity/rate of any computer much smaller than the universe or slower than the cosmological speed limit. But isn’t that really what Dr. Dembski’s UPB says? Its interesting to relate essentially the same results derived via different methods. I suspect that such results are really telling us something. I will upon demand relate many such results and specifically biological. Dr. Dembski does occasionally allude to such results. But these results are unnoticed by his critics. (Dr. Dembski needs, as the critics have said, to relate his methods and results to more "conventional" methods and results (that is how the results are related). I do not in any way want to discourage an original approach to the problem! I.e., not "original" but unfamiliar (to biologists).
Its also interesting to note that there is a life form that (annoyingly) violates the axioms of probability theory—humans. Yet humans, who consistently violate those axioms in their thinking, are also marvelously well-adapted, according to the theorists, because of their thinking! This is a paradox for any probability theory that asserts that it defines the laws of the universe, but admits that life forms, well-adapted to the universe, do not abide by those laws.
Basically the “laws of probability” are not the laws of life, and that is my criticism of Dr. Dembski. But I think he can turn that into a positive for his thesis. Whereas it is highly problematic for any thesis that insists that they must be the same, except, as I often hear here, where these laws cause us some embarrassment.
Also what receives little consideration, other than stark incredulity, is some of the numbers Dembski produces. Even if I accept these numbers as a “worst case” only, then any process that reduces these numbers significantly (and by significantly I do not mean to “save the phenomenon,” i.e. received theory), we are positing the existence of some naturally occurring “mechanism” that by far exceeds the “efficiency” of any known! (Or any possible such “mechanism” as dictated by the ever-confounding laws of probability-SLOT.) The only possible explanation here is a “law.” Yet, evolutionary theory rejects such a law! Even if it accepted such a law, it plays right into Dembski’s hands! Darwin referred to natural selection as a “law,” under his naïve conception, i.e., non-current conception of what “law” means for scientists: a description not a prescription. But what we are requiring, or what is “necessary” here is a prescription. It’s ironic to consider that we are (Unwittingly? In the way the phylogeneticists wittingly do?) referring evolution to a “fixed outcome.” That this “law” strictly determines such. But that is even implicit in Dembski’s UPB. Basically nothing that is really interesting, e.g., any life form, is the result of the “hard over contingency” of theory. It is purely lawful! It is the result of the “operation” of a “law”!
I can’t imagine that as scientists we are interested in how “God did it” or how “God didn’t do it.” But only in how it is done.
(Tangentially only, has anything of this to do with “specified complexity?” Fact is that biologists have long assumed (with good reason) that biology is specified and complex. If they lack a formal theory with which to work, than all credit is due to Dr. Dembski for attempting to make such a contribution. This is something that needed to be done. If it results in some really interesting questions and problems, and I can't imagine that anyone participating or contributing here doesn't think it is... an interesting problem... then I suppose that the usual criticism is false. Taking a design perspective is productive.)
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 12. May 2003 16:02
Nelson wrote:
quote:
If the specification is f(10) then I have F(1) through f(10) as specifications.
Exactly! At least, that is what I think should happen, assuming that you meant that the observed event was f(10). However, that is not what Dembski's mathematics in NFL gives you. That is what my correction gives.
Nelson also asked me to remove a letter and retain the meaning of:
"Can you show me how you remove just one of the letters in this sentence and still retain it's meaning?"
If he is speaking of its intended meaning, then I say yes:
"Can you show me how you remove just one o the letters in this sentence and still retain its meaning?"
"Can you show me how you remove just one of the leters in this sentence and still retain it's meaning?"
"Can you show me how you remove just one of the letters in this sentece and still retain it's meaning?"
"Can you show me how you remov just one of the letters in this sentence and still retain it's meaning?"
"Can you show me how you remove jus one of the letters in this sentence and still retain it's meaning?"
"Can you show me how you remove just one of the letters in this sentence and stil retain it's meaning?"
I presume everyone still understands all of these sentences, but I'm not sure what this is supposed to illustrate. We have defined a proper form for spelling and sentence construction, but we have error-correcting capabilities and thus small typos (like it's for its in the original) can be overcome. (Even MS Word can handle this level of error.)
I'm not sure how likely it was for those letters to appear on my screen, but I am equally unsure that they are specified. To the degree that they are unlikely, they seem also to be unspecified.
As such, a Pdco calculation alone isn't of much use. The determination of specification must also be done.
Dembski's best-worked-out example seems to be the Caputo case. I'll try to at least touch on the steps that have been skipped. It is helpful here to refer to NFL pp. 80-83, since I'm not going to retype the entire argument.
In step #2, we blithely conclude that "the only chance process that could have been operating to produce [the event] was one in which the Ds and Rs are equiprobable and probabilistically independent of one another." The reason is because Mr. Caputo said that this is what he did. That's fine, if we, like the Supreme Court, are only trying to figure out whether Mr. Caputo is lying or not. However, there are other processes that deserve more consideration, if we really want to rule out all chance hypotheses and regularities. For instance, perhaps Mr. Caputo took a shortcut and didn't use the Urn method like he was supposed to; aren't those relevant to the outcome? Humans regularly fix things to get desirable outcomes; perhaps Caputo did the same. If we parameterize human behavior with a probability density function, the outcomes of 41 election ballot order selections is going to look somewhat different from the binomial distribution generated by random selection. What is left out of #2 (and everywhere else I can find) is a detailed discussion of why only that one chance hypothesis was relevant, and how we can know that it is the only one. What steps do you take to make that determination?
In step #4, the assertion is made that the Supreme Court knows about counting and therefore can identify the rejection region independently of the event. However, it is just an assertion. What should be shown is that (1) background knowledge K explicitly and univocally identifies the rejection function f and region R; (2) P(E|H&K) = P(E|H). It is fairly clear that (2) holds for the H that Dembski has picked (i.e. a Bernoulli process with p=0.5), as long as K is independent of the event. However, that doesn't seem to be what's going on; even if we've identified a counting function f(E) = # of Democrats, based on our ability to count, why are we counting the Democrats and not the Republicans? And why do we set our rejection region at 40 of 41? That doesn't seem to be very independent of E.
In step #5, Dembski simply claims that he has picked simple specifications. Where is the objective complexity function phi (p. 76)? I agree that he has picked simple-looking specifications, but this is a critical step and therefore we should avoid simply relying on intuition if at all possible. In particular, I am worried that intuition may cause us to interpret specifications as simple or complex in the context of the event we've observed--not independent of it!--and thus underestimate SpecRes. In order to rule out this bias, a formal method of analysis is highly desirable.
Janitor, I agree with you that it may well be difficult to consider design and non-design as cleanly separable processes.
I am not sure what you mean when you say that "life is reflexive". What is the relation R, and what set X are elements drawn from? It's difficult to appreciate the problem without knowing what you mean.
Also, I no longer remember what a "strong Markov condition" is. Could you remind me, and briefly indicate how it is (or is not) relevant to biological and physical processes? Is it a problem if life forms only achieve partial and ad-hoc orderings?
IP: Logged
|
|
Erik
Member
Member # 160
|
posted 13. May 2003 15:20
Rex Kerr asked (in the context of the Caputo example): "And why do we set our rejection region at 40 of 41? That doesn't seem to be very independent of E."
Actually, Dembski's method does not require that the limit of the rejection region is set independently of the event. The requirement is that the knowledge used to identify the rejection function must be "independent" (in a sense that is still unclear) of the event E. Indeed, when using the rejection function to decide a specification of the form
S = {w : f(w) >= c}
Dembski has encouraged us to choose the largest value that is consistent with S being a superset of the observed event. In the Caputo case, this led to c = 40.
Therefore your concern is not a fair criticism of the specific example application of the EF, although it could be a fair criticism of the EF itself. (Unless you're a hardcore Bayesian, you may want to take care to avoid that the criticism of the choice of c will also apply to the choice of rejection region is classical hypothesis testing, though. Fisher's method asks us to reject the null hypothesis if, e.g., Pr(t(X) > t(x) | H0) is less than a critical value, where t is the test statistic, x the observed data and X is the corresponding stochastic variable.)
I agree with the rest in the post I quoted from, though.
Erik
IP: Logged
|
|
Nel
Member
Member # 614
|
posted 13. May 2003 15:26
Note, since I'm done with the topo thread, which contained many different topics that had to be addressed in a single post, I'm able to spend more time here responding to Erik and Rex, unfortunately I'll have to get to that tomorrow.
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 13. May 2003 23:42
As I have argued in other threads and perhaps a bit too often for the liking of the moderator, the ID inference is an eliminative approach.
Behe recently made some comments which seem to support my objections quite clearly and in the interest of the discussion I would like to present the opinion from the ID proponent's perspective:
quote:
The peril of negative arguments is that they may rest on our lack of knowledge, rather than on positive results.
Source
Behe made his comments in objection to the argument "an intelligent designer would not have done it this way" but I believe that his statement even better applies to the ID inference.
In fact ignorance seems to be a relevant and important source of false positives in the ID inference. Hence the contributions of Elsberry and Wilkins to improving the ID filter seem quite relevant.
[ 13. May 2003, 23:45: Message edited by: Pim van Meurs ]
IP: Logged
|
|
Rex Kerr
Member
Member # 632
|
posted 14. May 2003 14:32
Erik, I agree with your points; my post perhaps was fishing more than it should have been. I want to actually see the set of knowledge that is being used to generate the specification, since whenever I do it, it ends up using an apparently large amount of knowledge. This is a problem, because when you then come to #5, the complexity is high.
But perhaps I have the wrong idea of what "explicitly and univocally" means; it's really hard to tell until I see the knowledge and a demonstration of how it identifies the rejection region and function.
IP: Logged
|
|
gedanken
Member
Member # 594
|
posted 16. May 2003 00:20
quote:
Actually, Dembski's method does not require that the limit of the rejection region is set independently of the event. The requirement is that the knowledge used to identify the rejection function must be "independent" (in a sense that is still unclear) of the event E. Indeed, when using the rejection function to decide a specification of the form [...]
I'd looked at that as well, but decided that if it was a problem, it was repairable. By that I mean that one can easily fix Dembski's argument in thar regard if it is considered to be a problem.
Rather than set the threshold from the event, one simply asks if there exists a threshold (for the "rejection region" that contains the event, and which also has the probability below the cutoff probability. (I'm going by memory, so don't have exact terminology.)
This would be totally independent of the event, as any trial threshold would not depend on the event. Of course the answer to the question, does it meet the conditions, would depend on testing if the region contained the event -- obviously one can't avoid that. This is not the problem with the EF.
IP: Logged
|
|
Janitor@MIT
Member
Member # 125
|
posted 16. May 2003 14:57
With apologies to Mr. Moderator for the personal nature of this post; it’s been educational and entertaining, but its time I moved on. I can’t help but be disappointed that the “mystery of mysteries” that Whewell referred to hasn’t been answered for me. But, hey, if life was devoid of mysteries it wouldn’t really be worth living, would it? Hmmm… makes me think. LOL I appreciate that ya’ll made me think. Thinking seems so natural that we often forget just how difficult it is to actually do. I cannot express to you my admiration for the fact that you think. Keep it up! LOL Love ya’ll. Goodbye.
IP: Logged
|
|
Nel
Member
Member # 614
|
posted 11. June 2003 00:04
Sorry for the delay in replies in this thread, some important things I had to do came up, I'll catch up on this board evenutally. I'm going in chronological order. I also want to get involved in some of the EQU discussions which are similar, so I won't be dredging up too many old arguments.
Erik,
C10: It is simply not true that the set of hypothesis we consider does not "mean" the sample space. According to both conventional statistics and Dembski's use of it, the hypothesis we consider (all possible outcomes of an event) can be one and the same. For example, in a sample space all outcomes have to be possible, even the flagellum forming by chance alone. There also exists the possibility that natural selection built the flagellum step by step. However, we do not have any detailed hypothesis concerning this to take this possibility seriously. The problem, in other words, is that there is no plausible sample space, that is the gist of my response to C10, the one under consideration No Free Lunch, natural selection along a single axis, yielded a low probability value. I do not see how this calculation is ambiguous.
With respect to Dembski's standards this is incorrect:
quote:
1. The sample space.
2. The event under study.
3. The set of hypotheses that are considered.
4. The rejection function(s) used.
5. The background knowledge used to explicitly and univocally identify the rejection function(s).
Dembski is using Fisher's approach, which is a conventional statistical method. It looks more like this:
1) Probability space
2) Null hypothesis
3) The probability distribution defined on this space
4) The observed event.
5) The specification
You say that I have not managed to provide a detailed calculation to the effect of one observed outcome, however, I did provide it, the calculation for the bacterial flagellum. Saying that none exists without a response to that calculation is not enough. In my responses to Rex, I plan to discuss this a little more.
C11: As an aside, simply asking for a reference to one of my quotes is good enough. However, on to your reply, it didn't address any of my arguments other than to say that Davies never says that chance cannot give rise to specified complexity, but that is irrelevant. The definition of specification does not require one to believe that chance cannot possibily give rise to specificied complexity. In fact, Dembski himself states that specified complexity does not rule out the logical possibility of it forming by some freak accident, rather than intelligent design. The consequence of specified complexity is that it is unlikely for a natural process to accomplish it.
The definition of an external pattern is not that it could not have arisen by some non-ID process either, that is also quite false. The definition of an external pattern is that it is more than what it is made up of. For example, an irreducible set of parts is an external pattern, and yet I think that random chance alone, no natural selection needed, can accomplish an external pattern quite easily. Consider randomly banging on your keyboard and as you do so you have managed to write out the word "the". You have achieved specificity through a stochastic process, an external pattern that is not reducible to the individual letters.
C12: Your response to my reply to C12 was simply a "switcheroo", Biologists don't test their Darwinian hypothesis statistically, so there is no way to know whether natural selection and random mutation would likely generate something like the bacterial flagellum. You turned this around and said that the failure of Darwinists to test their hypothesis is irrelevant to Dembski's apparent failures, however, I am arguing that critics of Dembski's method are not pointing to the any real flaws, instead, they simply do not believe that statistical tests are relevant. However, I think that when faced with a competing theory like ID, and I won't really get into all the reasons why, I think Darwinists should take attempts to statistically test their hypothesis seriously.
C13: neo-Darwinism is based on the extrapolation from micro evolutionary processes streched to long time periods. Single step changes, where a beneficial change is selected for, is quite easily accomplished by intelligent agents. Whether we can recreate earthquakes or tornados is irrelevant in this context, however doing so is just a matter of effort and time. If we created the atom bomb, I don't see how we wouldn't be able to invent something that can cause earthquakes.
I don't know what the "DEF" stands for, and I don't know how it is relevant to the fact that Dembski's EF is similar to the scientific method's use of the null hypothesis. So I'm not sure how your criticism "stands".
C14: My reply to C14 was not irrelevant. My reply mentioned that the universal probability bound is a way to limit what background knowledge we can use in order to obtain a rejection function.
[ 11. June 2003, 00:05: Message edited by: Nelson_Alonso ]
IP: Logged
|
|
Pim van Meurs
Member
Member # 541
|
posted 11. June 2003 00:54
Nelson: However, I think that when faced with a competing theory like ID, and I won't really get into all the reasons why, I think Darwinists should take attempts to statistically test their hypothesis seriously.
Why? Darwinists unlike ID'ers need not rely on probability tests for their hypotheses. Especially when such tests are almost impossible to accurately perform. And that is the main problem for ID based on a design inference which tries to reject any and all competing regularity and chance pathways. In order to do so it has to perform calculations using data to which it does not have access.
Houston we have a problem...
There are countless problems with Dembski's thesis but in the end it all comes down to the inability to accurately calculate the necessary probabilities will inevitably block any design inference conclusion.
IP: Logged
|
|
|
Printer-friendly view of this topic